Shadi Iskander
@shadi_isk
Followers
36
Following
19
Media
0
Statuses
9
MSc student @ Technion | Passionate about NLP research
Haifa, Israel
Joined September 2022
Excited to share that our paper on data quality evaluation for tool-using LLMs has been accepted to EMNLP 2024 Main🎉 We propose two automatic data quality evaluation metric, showing that high-quality, smaller datasets outperform larger, unvalidated ones. https://t.co/N9sDrsJzSe
arxiv.org
Training large language models (LLMs) for external tool usage is a rapidly expanding field, with recent research focusing on generating synthetic data to address the shortage of available data....
1
0
4
Why it matters? - Ensures models trained for tool usage are more reliable. - Helps optimize resources, allowing for smaller, well-curated datasets instead of relying on massive, noisy data. #EMNLP2024
0
0
1
We show that smaller, high-quality datasets led to better or comparable performance than larger, unvalidated sets. Proven on ToolBench and ToolAlpaca benchmarks, highlighting the critical role of data quality in LLM training.
1
0
2
We assess synthetic datasets used for training LLMs on tool usage. Propose two data evaluation methods: 1. Human-defined criteria 2. In-Context Evaluation (ICE)
1
0
1
Attending #NAACL🇲🇽? Come check out our poster and learn how we tackle social bias in language models without explicit demographic information. Meet me tomorrow at 9:00, Poster Session 4. https://t.co/qiaR24g9xH
@boknilev @KiraRadinsky
arxiv.org
Mitigating social biases typically requires identifying the social groups associated with each data sample. In this paper, we present DAFair, a novel approach to address social bias in language...
0
5
20
Can you list three influential people from the 19th century? (Please answer without using ChatGPT or any other generative AI.) And please retweet. @MichalShur
15
13
11
- Paper: https://t.co/lki2IHLMDy - Presentation Schedule @aclmeeting : Thursday, July 13, 16:30pm EDT @ Harbour B #Repl4NLP Friday, July 14, 11:40am EDT @ Pier 4 @trustnlp Join us to learn more about our work! #ACL2023NLP #NLP #Fairness
0
0
3
Our research introduces a novel approach that effectively eliminates non-linear encoded concepts from neural representations, aiming to address social biases and enhance fairness in natural language processing models.
1
0
1
I will present at #ACL2023 our poster of the ACL Findings paper titled 'Shielded Representations: Protecting Sensitive Attributes Through Iterative Gradient-Based Projection' in collaboration with @boknilev and @KiraRadinsky.
1
4
9